Does Dataset Complexity Matters for Model

Explainers?
1st José Ribeiro 2nd Raı́ssa Silva 3nd Lucas Cardoso
Federal University of Pará - UFPA IRMB, Montpellier University ICEN, Federal University of Pará
Federal Institute of Pará - IFPA La Ligue Contre le Cancer Belém, Brazil
Belém, Brazil Montpellier, France lucas.cardoso@icen.ufpa.br
jose.ribeiro@ifpa.edu.br r.lorenna@gmail.com

4rd Ronnie Alves

arXiv:2107.02661v2 [cs.LG] 17 Nov 2021

Federal University of Pará - UFPA

Vale Institute of Technology - ITV DS
Belém, Brazil
ronnie.alves@itv.org

Abstract—Strategies based on Explainable Artificial Intelli-
gence - XAI have emerged in computing to promote a better
understanding of predictions made by black box models. Most
XAI measures used today explain these types of models, generat-
ing attribute rankings aimed at explaining the model, that is, the
analysis of Attribute Importance of Model. There is no consensus
on which XAI measure generates an overall explainability rank.
For this reason, several proposals for tools have emerged (Ciu,
Dalex, Eli5, Lofo, Shap and Skater). An experimental bench-
mark of explainable AI techniques capable of producing global
explainability ranks based on tabular data related to different
problems and ensemble models are presented herein. Seeking
to answer questions such as “Are the explanations generated
by the different measures the same, similar or different?” and
“How does data complexity play along model explainability?”
The results from the construction of 82 computational models
and 592 ranks shed some light on the other side of the problem
of explainability: dataset complexity!
Index Terms—Explainable Artificial Intelligence - XAI, Black
box model, Dataset complexity.
I. I NTRODUCTION
Recently, technology has increasingly evolved and allowed
intelligent algorithms to be present in our daily lives through
solutions to the most diversified types of problems, thus
prompting a need for machine learning models to solve
increasingly complex problems which require characteristics
that justify decision making, in addition to high problem-
solving performance [1] [2].
Computational models based on bagging and boosting
algorithms, which provide for high performance and high
generalization capacity, are commonly used in computation to
solve regression and classification problems based on tabular
data. However, these models are not considered transparent
algorithms1 , being considered black box algorithms2 and,
1 Transparent Algorithms: Algorithms that generate explanations for how a
given output was produced. Such examples include Decision Tree, Logistic
Regression and K-nearest Neighbors.
2 Black Box Algorithms: Machine learning algorithms that have classifica-
tion or regression decisions that are hidden from the user.
therefore, less used in problems related to sensitive contexts,
such as health and safety [3] [4] [5].
The limited understanding of black box models requires a
search for measures or tools that can provide information about
local explanations — aiming to predict around an instance
through various methods to obtain a local attribute importance
ranking [6] — and global explanations — when it is possible to
understand the logic of all instances of the model by generating
a global attribute importance rank [6] [7] — as a means of
making decisions interpretable and, thus, more reliable [8].
Efforts have been made in the Explainable Artificial In-
telligence - XAI area regarding the development of different
measures to explain black box models, even after their train-
ing and testing process, called post-hoc analysis [3]. Thus,
measures such as Ciu [9], Dalex [10], Eli5 [11], Lofo [12],
Shap [13], and Skater [14] emerged to promote the creation of
model-agnostic3 explanations. Each of these tools is capable of
generating explanations using different techniques, but a fact
they have in common is that they all generate global attribute
importance rankings related to the explanation of a model.
The proposal to analyze the ranking of global importance
— and not the local — allows general analysis of how a given
model treats and explains the problem to be solved, and for
this reason, it was chosen for this study.
This research raises discussions for the current moment in
the XAI area through two main questions, namely: — Con-
sidering the current measures geared at explaining black box
machine learning models, can it be inferred that they generate
global rankings of same, similar, or different explainabilities?
— Following the same idea as in the previous question, are
the generations of equal, similar, or different explainabilities
related to specific properties of a dataset?
Seeking to answer the two hypotheses above, this research
emerges as a comparative analysis of different XAI measures,
3 Model-Agnostic: means it does not depend on the type of model to be
explained.
capable of producing global explainability ranks based on tab-
ular data related to different problems and ensemble models.
The main contributions of this research to the machine
learning area focused on XAI are as follows:
• Development of a benchmark that measures the corre-
lations existing between the explanations generated by
the main current XAI measures by comparing them with
each other through different perspectives (properties of
the datasets used);
• Generation of comparative results between explainability
ranks that lead researchers in the XAI area to identify
which explainability measures to use against different
characteristics of a dataset.
II. BACKGROUND
A. Explainable Artificial Intelligence
Recently, there is a growing need to explain models of
black box machine learning in an agnostic way. This includes
not only more robust computational models like Deep Neural
Networks but also simpler models like ensembles [15] [16]
[17] [18].
In this way, the herein research identifies in the current
literature the necessity of specific studies on model-agnostic
XAI measures for tabular data, thus enabling explanations
of computational models with wide applicability, such as
ensemble trees [17] [3].
A bibliographic and practical survey (development) on the
main existing XAI measures was performed through this re-
search, specifically aimed at generating model-agnostic global
explainability ranks that support tabular data. As a result, a
total of six tools were selected, with updated and compatible li-
braries with current Python development platforms and Scikit-
Learning algorithms [19]. These tools are: CIU [9], Dalex [20],
Eli5 [11], Lofo [12], SHAP [13] and Skater [14].
Note that all the six measures referenced herein generate
explanatory ranks based on the same machine learning models
previously trained, manipulate their inputs or/and produce new
intermediate models (copies). Therefore, a comparison of the
generated explanatory ranks is fair and feasible.
Due to the incompatibility between the current XAI mea-
surement libraries and the versions of the scikit-lean [19]
computational model libraries and dependencies, only the
Random Forest and Gradient Boosting algorithms were used
in this study.
During the initial analyzes and executions, this research
found XAI measures with incompatibilities at the level of
libraries and dependencies, which made it impossible to create
uniform tests with the other measures. For this reason, some
XAI means were not used for this research (for example:
Alibe-ALE [21], Lime [22], Ethical ExplainableAI [23], IBM
Explainable AI 360 [24], and Interpreter ML [25]).
Contextual Importance and Utility - CIU is a XAI measure
based on Decision Theory [26] that focuses on serving as a
unified measure of model-agnostic explainability based on the
implementation of two different indexes, namely: Contextual
Importance - CI and Contextual Utility - CU, which generally
create equal attribute ranks, but with different values for each
index, thus generating global explanations [9]. Only the CI is
used in this research, for obeying the same CU rank sequence.
Dalex is a set of XAI tools based on the LOCO (leave-one
covariate out) approach and it can generate explainabilities
from this approach. In general, the measure receives the model
and the data to be explained; it calculates the performance
of the model, performs new training processes with new
generated data sets, and it does the inversion of each attribute
of the data in a unitary and iterative way, thereby measuring
which attributes are most important to the model based on the
performance [27] [10].
Leave One Feature Out - Lofo is a XAI measure that has
a similar proposal than Dalex, but with a main difference
regarding the inversion of the attribute iterativity, because in
the Lofo measure the iterative step is the removal of the
attribute to find its global importance to the model based on
the performance [12].
Explain Like I’m Five - Eli5 is a tool that helps explore
machine learning classifiers and explains their predictions by
assigning weights to decisions, as well as exporting decision
trees and presenting the importance of the attributes of the
model submitted to the tool [11].
SHapley Additive exPlanations - SHAP is proposed as
a unified measure of attribute importance that explains the
prediction of an instance X from the contribution of an
attribute. The contribution of this attribute is calculated from
the game theory of Shapley Value [28]. In this tool, it is
possible to calculate the average of the Shapley Value values
by attributes to obtain the global importance [13] [29].
Skater is a set of tools capable of generating rankings
of the importance of model attributes, based on Information
Theory [30], through measurements of entropy in changing
predictions, through a perturbation of a certain attribute. The
central idea is that the more a model is dependable upon an
attribute, the greater the change in predictions [14].
A general comparison of the main characteristics of the
measures that meet the requirements of this research is pre-
sented in table I.
All the tools described above are capable of generating
explainabilities that go beyond the generation of attribute
importance ranks; therefore, the focus of this article is to
compare this most basic unit of model explainability, i.e., the
global importance rank.
With the explanatory ranks generated by the different mea-
sures, it is fair to compare existing correlations between these
ranks, even if each measure generates them based on different
algorithms, since the central objective of each measure is to
explain a black box model.
B. Dataset Properties
After being processed through attribute engineering, a
dataset is expected to exhibit characteristics that involve the
nature of the problem in a contextual (problem properties) and
technical (dataset properties) manner and which are general-
 TABLE I
M AIN XAI MEASURES SURVEYED

Global
XAI Local API
Autor Algorithm basis explanation
meansure explanation compatible
(by rank)
CIU [9] Decision Theory Yes No Yes
Dalex [20] Leave-one covariate out Yes Yes Yes
Eli5 [11] Assigning weights to decisions Yes Yes Yes
Lofo [12] Leave One Feature Out Yes No Yes
SHAP/Tree [13] Game Theory Yes Yes Yes
Skater [14] Information Theory Yes Yes Yes

ized by machine learning algorithms to carry out a prediction
or a classification process [8].
Although the contextual side of the problem significantly
helps in interpreting the explainabilities of the models, the
herein researchers opted not to work with these characteristics
of the analyzed data, since datasets utilized refer to differ-
ent problems with contexts that often go beyond computing
knowledge.
On the other hand, there are properties of datasets which are
feasible to be analyzed, since they are common — even though
these datasets represent different problems. Properties such
as dimensionality, number of numerical attributes, number of
binary attributes, the balance between classes, and entropy are
examples of properties that every tabular dataset has and which
directly influence the generalizability of the proposed model
[31].
By the time this article was published, research that con-
fronted the inherent characteristics of different datasets (com-
plexities) and their explanations through machine learning
models was not identified in the literature. In this sense, this
research seeks to fill in this gap in the framework of XAI
measure studies.
III. M ATERIALS AND M ETHODS
This research developed a benchmark, Figure 1, which uses
tabular data already known and endorsed by the machine
learning community (Figure 1 - 1), extracted properties from
these datasets (Figure 1 - 2), clustered these datasets according
to their properties (Figure 1 - 3), performed the construction
of computational models (Figure 1 - 4 and 5), generated
explainability ranks for all models (Figure 1 - 6 and 7),
calculated the correlations obtained from the all rank pairs and
analyzed the different results (Figure 1 - 8 and 9). All these
procedures are explained in detail in the following topics.
A. Datasets and Preprocess
A total of 41 datasets of binary problems were used, with no
data loss, and selected from the OpenML [32] platform. The
most used datasets were selected, so as to make the results

Fig. 1. Visual scheme of all steps and processes performed by the proposed benchmark.
of this research even better for use by the machine learning
community.
The datasets used were as follows: australian, phishing
websites, spec, satellite, analcatdata lawsuit, banknote authen-
tication, blood transfusion service center, churn, climate model
simulation crashes, credit-g, delta ailerons, diabetes, eeg-
eye-state, haberman, heart-statlog, ilpd, ionosphere, jEdit-4.0-
4.2, kc1, kc2, kc3, kr-vs-kp, mc1, monks-problems-1, monks-
problems-2, monks-problems-3, mozilla4, mw1, ozone-level-
8hr, pc1, pc2, pc3, pc4, phoneme, prnn crabs, qsar-biodeg,
sonar, spambase, steel-plates-fault, tic-tac-toe and wdbc.
All datasets went through the following pre-processing
steps, as required: converting categorical attributes to
frequency-based ordinals, converting boolean attributes to in-
teger (0 and 1), and min-max normalization to values between
0 and 1 [33].
B. Clustering and Multiple Correspondence Analysis
Analyzing different datasets based on their properties allows
for comparing similarities and differences between them, even
if these datasets are from different contexts. For example, we
can group a set of datasets based on their class entropy values
and then identify which datasets have more and less informa-
tion, thus enabling a dataset complexity impact relationship
for future model explanations.
Based on the proposal laid out above, this research used
Fig. 2. Silhouette coefficients for clustering, using the Kmeans algorithm, for
15 different properties (provided by OpenML) extracted from a = 3. Distance means (x axis) and label of clusters 0, 1 and 2 (axis y).
each of the 41 selected datasets and then used the clustering
algorithm k-means [34] to identify groups of datasets based
on their similarities. categories of nominal variables (p) — here are the properties
The interpretation and validation algorithm between data of each dataset.
clusters, called Silhouettes [35], was used, the K value (clus- This research performed the process of binarization in the
ters) of which ranging between 2 and 20, and K = 3 with same table in which clustering was applied, replaced by h
Average Silhouette Scores = 0.28 being identified, Figure 2. (equal to or above the mean of attribute values) and s (below
In figure 2, it can be seen that for a k=3 the distances the mean of attribute values). For more information about the
between each cluster (0, 1 and 2) are above the average (red new properties dataset, refer to: https://github.com/josesousa
line), so this is an adequate value of k in relation to the study ribeiro/XAI-Benchmark/blob/main/Openml/df properties bin
presented hereby. arized.csv.
The following 15 properties being used included: Number This analysis allows for identifying different relationships
of Features, Number of Instances, Dimensionality, Percentage between dataset (or even dataset group) and properties (or even
of Binary Features, Standard Deviation Nominal of Attribute property value ranges), as advocated by the literature [36].
Distinct Values, Mean Nominal Attribute Distinct Values, Class
Entropy, Autocorrelation, Number of Numeric Features, Num- C. Construction of Models
ber of Symbolic Features, Number of Binary Features, Percent- The algorithms used in building each dataset model were
age of Symbolic Features, Percentage of Numeric Features, Random Forest - RF and Gradient Boosting - GB, both being
Majority Class Percentage, and Minority Class Percentage. tree-based, low-explainability ensembles [3].
Further information about the values of each property for This research tested newer ensemble algorithms such as
each dataset can be found in: https://github.com/josesousaribe Light Gradient Boost [37], CatBoost [38], and Extreme Gra-
iro/XAI-Benchmark/blob/main/Openml/df dataset properties. dient Boosting [39]; however, not all XAI metrics that were
csv. used supported the output encodings of these new models.
In order to identify the different correlations between each Thus, an option was made to adhere to the Gradient Boosting
dataset cluster identified with the 15 properties listed above, and Random Forest algorithms, only.
this research used the Multiple Correspondence Analysis - Notably, the Tuning step was performed through Grid
MCA [36] in the properties dataset, where the rows of this Search [40] with cross-validation [41] of f old = 3, thus that
table are the observations or individuals (n) concerned — the algorithms had a better calibration on the analyzed data,
here are the datasets — and the columns are the different and a total of 192 candidates through the parameters: max
depth: (1, 10), bootstrap: (True, False), n estimators: (100,
200), min samples leaf: (1, 10), min samples split:(2,10), max
features: (sqrt, log2), and loss: (deviance, exponential) (the
latter only valid for the GB algorithm).
Finally, each model was trained and tested respectively
based on a 70% − 30% ration for each dataset. Then, each
model was measured as to their performance and stability
regarding accuracy and precision, recall and Friedman test
[42], so as to characterize each model that was created.
D. XAI Ranks and Correlations
For each of the 41 datasets (divided into 3 different clusters),
2 models were created, one based on the Random Forest al-
gorithm and the other on Gradient Boosting, thus generating a
total of 82 models. Global explainability ranks were produced
through the 6 different XAI measures for each model, resulting
in a total of 492 ranks.
The Spearman Rank Correlation [43] was used to calculate
the correlations between the ranks that were created. The
reason for using this algorithm in particular, is that it measures
the correlation between rank pairs considering the idea of
ranks (positions) in which different values (in this case, dataset
attributes) may appear. In this step, two comparison matrices
of rank correlation pairs are generated for each dataset (one
matrix for each algorithm). Figure 3 shows an example of this
matrix created from the Gradient Boosting - GB model.
Fig. 3. Example Spearman correlation matrix (represented in value and color
in each cell) based on explicability ranks generated from Gradient Boosting
model to the dataset wdbc. Note that, for this dataset, chosen at random, the
correlations between the rank pairs are low or non-existent (close to zero).
Correlations such as those shown in Figure 3 are calculated
for each dataset and show how the correlations between
different pairs of XAI measures are presented for each problem
and for each computational model being developed.
The results shown in Figure 3 refer to only one dataset
and, therefore, are intermediate results, since they will be
summarized in the results of the other datasets.
E. Benchmark of XAI measures
All the methodology described so far was consolidated and
implemented in a single benchmark. More details about the
implementation and execution of the procedures presented
hereby are available in the repository: — https://github.com/j
osesousaribeiro/XAI-Benchmark.
IV. R ESULTS AND D ISCUSSION
The results obtained were computed in groups (2 different
computational models) and clusters (3 different dataset clus-
ters).
The principle for understanding the results of this article is
to acknowledge the different dataset cluster profiles identified
in the clustering process and MCA.
In this sense, according to clustering, based on the 15
different properties of the 41 datasets that were analyzed, it is
possible to assess the existence of at least 3 different dataset
clusters. In quantitative terms, each cluster has 21 (cluster 0),
17 (cluster 1), and 3 (cluster 2) datasets.
To identify each different dataset profile, an inspection of
the value ranges of each cluster for each property was carried
out. This analysis verified the possible existence of specific
property values for datasets of specific clusters.
At this stage, it was identified that the quantity of datasets
in cluster 2 had a small n and, therefore, it was disregarded
from these results (the datasets for this cluster are: kr-vs-kp,
Phishing Web sites, and SPECT). Keeping only clusters 0 and
1.
Intending to consolidate the profile analysis of the clusters,
Multiple Correspondence Analysis - MCA was applied, and
the relation between the datasets and the value ranges of their
properties was verified, thus providing for a better understand-
ing of the complexity of the datasets existing in both clusters
— this practice is already known in the literature [36].
By inspecting the MCA shown in Figure 4, the relation
(shorter distances) between high quantity of datasets in cluster
0 and above-average values (symbol h) for properties (big-
ger circles) such as MajorityClassPercentage, Dimensionality,
NumberOfInstances, NumberOfFeatures, PercentageOfNumer-
icFeatures, and AutoCorrelation was identified. Also, a rela-
tion between a high quantity of datasets of cluster 0 and below-
average values (symbol s) for properties such as ClassEntropy,
PercentageOfSymbolicFeatures, and PercentageOfBinaryFea-
tures was observed.
Also with regard to the MCA shown in Figure 4, by inspec-
tion, assessing the existence of a relation (shorter distances)
between a high quantity of datasets in cluster 1 and above-
average values (symbol h) is possible for properties (big-
ger circles) such as ClassEntropy, PercentageOfSymbolocFea-
tures, PercentageOfBinaryFeatures, NumberOfBinaryFeatures,
and MinorityClassPercentage. It is also possible to identify
a relation between a high quantity of datasets in cluster
1 and below-average values (symbol s) for properties sush
as AutoCorrelation,NumberOfFeatures, PercentageOfNumer-
icFeatures, NumberOfNumericFeatures, NumberOfInstances,
Dimensionality, and MajorityClassPercentage.
 Fig. 4. Multiple Correspondence Analysis - MCA with rows (datasets) and columns (properties) depending on the first and second PCA components.
It is worth noting that some properties were not mentioned
above because they appear at very close distances for the two
dataset clusters.
However, based on the inspection of MCA results, it can be
concluded that cluster 0 is formed by a significant amount of
datasets with greater complexities, whereas cluster 1 is formed
by a significant amount of datasets with minor complexities.
The entire cluster profile identification process can be found
in: https://github.com/josesousaribeiro/XAI-Benchmark/blob
/main/Cluster Profile.md.
Based on the profile characterization of the two main clus-
ters that were identified, this research executed the Benchmark
for these two sets of datasets separately.
The benchmark execution for the datasets belonging to
cluster 0, Figure 5, exhibits low or negligible correlations
found in most tests, thus showing that the six XAI measures
used hereby generate different explainability ranks for most of
the datasets in this group, regardless of the algorithm (RF or
GB) the computational model is based upon.
The results presented by cluster 0, Figure 5, show that
Fig. 5. Boxplot graphs that summarize all rank pair comparisons (x axis) and their respective Spearman correlations (y axis) calculated for the datasets of
cluster 0. Results of the model based on Random Forest (left) and Gradient Boosting (right). The dashed blue lines show the different levels of correlations.
The points identify the positions of the correlation values of each comparison, and the color of the points refer to the attribute quantities of each dataset. Note
the low variance of most boxplots along with levels of negligent or low correlations found.

Fig. 6. Boxplot graphs that summarize all rank pair comparisons (x axis) and their respective Spearman correlations (y axis) calculated for the datasets of
cluster 1. Results of the model based on Random Forest (left) and Gradient Boosting (right). The dashed blue lines show the different levels of correlations.
The points identify the positions of the correlation values of each comparison, and the color of the points refer to the attribute quantities of each dataset. Note
the high variances of most boxplots along with the high levels of correlations found.
datasets with greater complexity generate more complex mod-
els, and in turn, different explanations. This could justify the
low correlations found in the results.
The benchmark execution for the datasets of cluster 1,
Figure 6, shows different correlations from those shown in
Figure 5, since there is high variance in boxplots as well as
high positive and negative correlations between rank pairs in
cluster 1.
The results presented in Figure 5 and 6 provide evidence
on how to answer the first hypothesis launched herein: “—
Considering the current measures aimed at explaining black
box machine learning models, can it be inferred that they
generate global rankings of same, similar, or different ex-
plainabilities?”. It depends, since the different correlations
found demonstrate that the properties of the datasets directly
influence the explainabilities of the models created from them.
Thus, it is possible to make the tools generate explainabilities
with higher correlations with one another (more similar ex-
plainabilities) and lower correlations with one another (less
similar explainabilities).
Noticeably, despite the similar results of the models based
on the RF and GB algorithms, there are important differences
in the correlations found, as shown in Figure 5. For example,
the larger variance of some comparisons in the results of the
GB algorithm (shap vs lofo, eli5 vs ci, eli5 vs skater, ci vs
skater and shap vs ci). This proves that the complexity of
the algorithm used in the model (in this case, bagging and
boosting tree-based) does have an influence on explainability
generation.
Importantly, the outliers existing in the boxplots of Figures
5 and 6, show the possibility of other dataset clusters existing
amongst those being analyzed. The colors, used to distinguish
the amount of attributes from different datasets in each boxplot
show that the results obtained in the correlations are not a
function of the number of attributes that a dataset has.
The comparison of the results presented in Figures 5 and
6 demonstrate the answer to the second hypothesis of this
research, namely: “— According to the same idea as in the
previous question, are the generations of equal, similar, or
different explainabilities related to specific properties of a
dataset?”. Yes, since the results of the rank correlations for
the two dataset clusters presented in the benchmark output
were different, thus showing that the complexity properties of
the dataset interfere with explainability ranks.
Based on the experiments performed with the 6 different
XAI measures and the resulted obtained therefrom, the follow-
ing logical reasoning could be constructed: 1 - If a ensemble
model (algorithm and dataset) solves a classification problem
for a low-complexity dataset, there must be few ranks (or even
only one rank) of explainabilities referring to this model, thus
allowing to infer that the different XAI measures have higher
correlations with each other (as seen in Figure 6); 2 - If an
ensemble model (algorithm and dataset) solves a classification
problem for a high-complexity dataset, there must be many
explainability ranks for this model, leading the ranks of the
different XAI measures to show lower correlations with each
other (as seen in Figure 5).
Based on the logical reasoning 1 and 2 above, there was
identified existence of evidence about Ensemble-Tree Models
with Complex Explainabilities - EMCX (created from the
data of cluster 0) and Ensemble-Tree Models with Simple
Explainabilities - EMSX (created from the data of cluster 1),
both terms proposed hereby. Furthermore, it was identified
that the divergence between the ranks is inherent, primarily
from the complexity of the dataset and, consequently, from the
generated model — thus answering the article’s title question.
The datasets depicted below, which served as the basis for
the creation of EMSX and EMCX, are intended to comply
with one of the contributions hereto:
• EMSX Datasets: Ozone level 8hr, Sonar, Spambase, Qsar
biodeg, Kc3, Mc1, Pc3, Mw1, Pc4, Satellite, Pc2, Steel
plates fault, Kc2, Pc1, Kc1, Climate model simulation
crashes, and Analcatdata lawsuit;
• EMCX Datasets: Ionosphere, Wdbc, Credit-g, Churn,
Australian, Eeg eye state, Heart statlog, Ilpd, Tic tac toe,
JEdit 4.0 4.2, Diabetes, Prnn crabs, Monks problems 1,
Monks-problems 3, Monks problems 2, Delta ailerons,
Mozilla4, Phoneme, Blood transfusion service center,
Banknote authentication, and Haberman;
It is clear that evidence of the existence of EMSX and
EMCX was observed from experiments with ensemble algo-
rithms. For these terms to become generally applicable (other
algorithms), experiments with algorithms of other natures are
necessary.
V. F INAL C ONSIDERATIONS
Given all the analyzes that were carried out, this research
achieves its objective of being able to observe the impacts of
dataset complexity in explaining black box machine learning
models, through the different properties of the data under anal-
ysis. That allowed for finding that the properties of a dataset
— facing the creation of models based on Random Forest and
Gradiant Boosting algorithms — show evidence that EMSX
and EMCX do exist. However, many other properties of the
data and of other algorithms still need to be further explored.
R EFERENCES
[1] S. Shalev-Shwartz and S. Ben-David, Understanding machine learning:
From theory to algorithms. Cambridge university press, 2014.
[2] Z. Ghahramani, “Probabilistic machine learning and artificial intelli-
gence,” Nature, vol. 521, no. 7553, pp. 452–459, 2015.
[3] A. Barredo Arrieta, N. Dı́az-Rodrı́guez, J. Del Ser, A. Bennetot,
S. Tabik, A. Barbado, S. Garcia, S. Gil-Lopez, D. Molina, R. Benjamins,
R. Chatila, and F. Herrera, “Explainable Artificial Intelligence (XAI):
Concepts, taxonomies, opportunities and challenges toward responsible
AI,” Information Fusion, vol. 58, pp. 82–115, Jun. 2020.
[4] R. Nambiar, R. Bhardwaj, A. Sethi, and R. Vargheese, “A look at
challenges and opportunities of big data analytics in healthcare,” in 2013
IEEE International Conference on Big Data, 2013, pp. 17–22.
[5] S. Lee, L. Chen, S. Duan, S. Chinthavali, M. Shankar, and B. A. Prakash,
“Urban-net: A network-based infrastructure monitoring and analysis
system for emergency management and public safety,” in 2016 IEEE
International Conference on Big Data (Big Data), 2016, pp. 2600–2609.
 [6] S. M. Lundberg, G. Erion, H. Chen, A. DeGrave, J. M. Prutkin, and Fairness in Python,” arXiv:2012.14406, 2020. [Online]. Available:
B. Nair, R. Katz, J. Himmelfarb, N. Bansal, and S.-I. Lee, “From local https://arxiv.org/abs/2012.14406
explanations to global understanding with explainable AI for trees,” [28] A. E. Roth, The Shapley value: essays in honor of Lloyd S. Shapley.
Nature Machine Intelligence, vol. 2, no. 1, pp. 56–67, Jan. 2020, Cambridge University Press, 1988.
number: 1 Publisher: Nature Publishing Group. [Online]. Available: [29] S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model
https://www.nature.com/articles/s42256-019-0138-9 predictions,” in Proceedings of the 31st international conference on
[7] R. Guidotti, A. Monreale, S. Ruggieri, F. Turini, F. Giannotti, and neural information processing systems, 2017, pp. 4768–4777.
D. Pedreschi, “A survey of methods for explaining black box models,” [30] F. M. Reza, An Introduction to Information Theory. Courier Corpora-
ACM computing surveys (CSUR), vol. 51, no. 5, pp. 1–42, 2018. tion, Jan. 1994, google-Books-ID: RtzpRAiX6OgC.
[8] D. Gunning and D. Aha, “DARPA’s Explainable Artificial Intelligence [31] D. Oreski, S. Oreski, and B. Klicek, “Effects of dataset characteristics
(XAI) Program,” AI Magazine, vol. 40, no. 2, pp. 44–58, Jun. 2019, on the performance of feature selection techniques,” Applied Soft Com-
number: 2. [Online]. Available: https://ojs.aaai.org/index.php/aimagazin puting, vol. 52, pp. 109–119, 2017.
e/article/view/2850 [32] OpenML, https://www.openml.org/search?q=qualities.NumberOfClass
[9] K. Främling, “Decision theory meets explainable ai,” in International es%3A2%2520qualities.NumberOfMissingValues%3A0&type=data&s
Workshop on Explainable, Transparent Autonomous Agents and Multi- ort=runs&order=desc, Last accessed 01 Mar 2021.
Agent Systems. Springer, 2020, pp. 57–74. [33] “MinMaxScaler,” https://scikit-learn.org/stable/modules/generated/
[10] P. B. a. T. Burzykowski, Explanatory Model Analysis. [Online]. sklearn.preprocessing.MinMaxScaler.html, Last accessed 15 Nov 2020.
Available: https://ema.drwhy.ai/ [34] “KMeans,” https://scikit-learn.org/stable/modules/generated/
[11] M. Korobov and K. Lopuhin, “Eli5,” sklearn.cluster.KMeans.html, Last accessed 2 Mar 2021.
https://eli5.readthedocs.io/en/latest/index.html, Last accessed 21 Jan [35] P. J. Rousseeuw, “Silhouettes: a graphical aid to the interpretation and
2021. validation of cluster analysis,” Journal of computational and applied
[12] U. Çayır, I. Yenidoğan, and H. Dağ, “Use case study: Data science mathematics, vol. 20, pp. 53–65, 1987.
application for microsoft malware prediction competition on kaggle,” [36] H. Abdi and D. Valentin, “Multiple correspondence analysis,” Encyclo-
Proceedings Book, p. 98, 2019. pedia of measurement and statistics, vol. 2, no. 4, pp. 651–657, 2007.
[13] S. M. Lundberg, G. Erion, H. Chen, A. DeGrave, J. M. Prutkin, [37] Microsoft, “LightGBM,” https://lightgbm.readthedocs.io/en/latest/, Last
B. Nair, R. Katz, J. Himmelfarb, N. Bansal, and S.-I. Lee, “From accessed 2 Mar 2021.
local explanations to global understanding with explainable ai for trees,” [38] Yandex, “CatBoost,” https://catboost.ai, Last accessed 5 Mar 2021.
Nature Machine Intelligence, vol. 2, no. 1, pp. 2522–5839, 2020. [39] “Xgboost,” Last accessed 5 Mar 2021. [Online]. Available:
[14] “Skater,” https://oracle.github.io/Skater/overview.html#skater, Last https://github.com/dmlc/xgboost
accessed 21 Jan 2021. [40] “Grid-search-CV,” https://scikit-learn.org/stable/modules/generated/skle
[15] Z. Qi, S. Khorram, and F. Li, “Visualizing deep networks by optimizing arn.model selection.GridSearchCV.html, Last accessed 10 Mar 2021.
with integrated gradients.” in CVPR Workshops, vol. 2, 2019. [41] “Cross Validation,” https://scikit-learn.org/stable/modules/cross validat
[16] P.-J. Kindermans, K. T. Schütt, M. Alber, K.-R. Müller, D. Erhan, ion.html, Last accessed 01 Mar 2021.
B. Kim, and S. Dähne, “Learning how to explain neural networks: [42] Friedman Test, https://scikit- posthocs.readthedocs.io/en/latest/gen
Patternnet and patternattribution,” in International Conference on erated/scikit posthocs.posthoc nemenyi friedman/, 2020, acessado:
Learning Representations, 2018. [Online]. Available: https://openrevi 02/07/2020.
ew.net/forum?id=Hkn7CBaTW [43] R. Artusi, P. Verderio, and E. Marubini, “Bravais-Pearson and Spearman
[17] P. Linardatos, V. Papastefanopoulos, and S. Kotsiantis, “Explainable Correlation Coefficients: Meaning, Test of Hypothesis and Confidence
ai: A review of machine learning interpretability methods,” Entropy, Interval,” The International Journal of Biological Markers, vol. 17,
vol. 23, no. 1, 2021. [Online]. Available: https://www.mdpi.com/1099- no. 2, pp. 148–151, Apr. 2002, publisher: SAGE Publications Ltd STM.
4300/23/1/18 [Online]. Available: https://doi.org/10.1177/172460080201700213
[18] “Ethical ML,” https://github.com/EthicalML/awesome-production-
machine-learning#explaining-black-box-models-and-datasets, Last
accessed 20 Apr 2021.
[19] “Scikit-learn,” https://scikit-learn.org/0.22/, Last accessed 21 Nov 2020.
[20] P. Biecek, “Dalex: Explainers for complex predictive models in r,”
Journal of Machine Learning Research, vol. 19, no. 84, pp. 1–5, 2018.
[Online]. Available: https://jmlr.org/papers/v19/18-416.html
[21] D. W. Apley and J. Zhu, “Visualizing the effects of predictor variables in
black box supervised learning models,” Journal of the Royal Statistical
Society: Series B (Statistical Methodology), vol. 82, no. 4, pp. 1059–
1086, 2020.
[22] M. T. Ribeiro, S. Singh, and C. Guestrin, “”why should I trust you?”:
Explaining the predictions of any classifier,” in Proceedings of the 22nd
ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, San Francisco, CA, USA, August 13-17, 2016, 2016, pp.
1135–1144.
[23] T. I. for Ethical AI and M. Learning, “Ethical XAI,” 2021, Last
accessed 22 Jul 2021. [Online]. Available:
https://ethical.institute/xai.html
[24] V. Arya, R. K. Bellamy, P.-Y. Chen, A. Dhurandhar, M. Hind, S. C.
Hoffman, S. Houde, Q. V. Liao, R. Luss, A. Mojsilovic et al., “Ai
explainability 360: An extensible toolkit for understanding data and
machine learning models.” J. Mach. Learn. Res., vol. 21, no. 130, pp.
1–6, 2020.
[25] H. Nori, S. Jenkins, P. Koch, and R. Caruana, “Interpretml: A uni-
fied framework for machine learning interpretability,” arXiv preprint
arXiv:1909.09223, 2019.
[26] R. L. Keeney, K. R. L, and R. Howard, Decisions with Multiple
Objectives: Preferences and Value Trade-Offs, revised ed. edição ed.
Cambridge England ; New York, NY, USA: Cambridge University Press,
Aug. 1993.
[27] H. Baniecki, W. Kretowicz, P. Piatyszek, J. Wisniewski, and P. Biecek,
“dalex: Responsible Machine Learning with Interactive Explainability